The escalating cost and protracted timelines of traditional drug discovery have spurred the adoption of computational strategies, especially those leveraging artificial intelligence (AI). In particular, deep learning (DL)-based de novo design has shown promise in generating novel drug-like molecules. However, rigorously validating these AI-generated compounds remains a significant hurdle. To address this challenge, we present an integrative workflow that combines molecular docking and molecular dynamics (MD) simulations for the systematic evaluation of deep learning-generated ligands targeting two pharmacologically critical proteins: the Adenosine A \(_{2}\) A Receptor (A2aR) and Ubiquitin-Specific Protease 7 (USP7). We first benchmarked six docking tools (AutoDock 4, AutoDock FR, AutoDock Vina, LeDock, PLANTS, and rDock) against reference ligands to identify those with the highest predictive accuracy. After selecting top-performing tools, we screened AI-generated compounds and applied exponential consensus scoring to refine hit prioritization. Finally, the most promising candidate complexes were subjected to all-atom MD simulations to assess binding stability and interaction fidelity. Our results indicate that AutoDock FR and AutoDock Vina consistently outperform other software in pose prediction, and that consensus scoring substantially improves hit enrichment. By providing a robust, multi-step validation pipeline, this study offers an efficient means to benchmark AI-generated molecules, ultimately facilitating more reliable drug candidate selection in both early-stage discovery and repurposing efforts.